deepseek-harness/docs/architecture.md

DeepSeek Harness Architecture

English | 中文

Read this before changing anything under packages/. It assumes you know Cordis; if you do not, start with the primer or the tutorial.

We recommend using an agent to explore the codebase and understand its architecture.

Cordis

Cordis is the framework under dsh: plugins contribute services, typed events, and reversible effects to a shared context. Every part of the product is a plugin, including the model adapter, the tool registry, the session log, and the agent loop itself, so each is replaceable from configuration.

There is no privileged core to patch: you extend dsh by mounting a plugin beside the others, and registrations are effects that unwind when their plugin unloads.

Profiles and bundles

A running dsh is a plugin tree composed at boot from ordered layers.

A profile is a named composition stored in the Harness home. It lists the bundles it stacks, holds any out-of-tree plugins it installs, and keeps the user's own cordis.patch.yml. web, headless, sdk, sdk-minimal, and acp ship as templates.

A bundle is a distribution format for Cordis config rows and the code they mount, so whatever it inserts stays patchable by the layers above it.

Each declares itself in its own package.json under a dsh field: dsh.profile lists a profile's bundles, and dsh.bundle points at a bundle's patch file.

dsh-base is the shared first layer of the web, headless, sdk, and acp profiles: model adapters, tools, persistence, sandbox and approval policy, settings, credentials, telemetry. dsh-web-app adds the browser application, dsh-headless adds a one-shot runner with no server, dsh-sdk-app adds the SDK JSON-RPC server, and dsh-acp-app adds the automation-only ACP server. dsh-sdk-minimal is the deliberate exception: one bundle owns its complete explicit SDK tree and does not apply dsh-base.

Layers apply to an empty entry list in this order: each bundle in the profile's listed order, then the profile's cordis.patch.yml, then the home-level one, then any --patch overlay. A patch targets a row by id and replaces its whole config, or inserts new rows.

Custom profiles default to live patch reload. The shipped web profile is live; headless, sdk, sdk-minimal, and acp apply all layers once at startup because replacing a one-shot or stdio application's dependencies after it owns work would invalidate that lifecycle.

To see the tree your machine boots:

dsh --profile web --dump-config

Any row it prints can be replaced by a patch of your own.

Composition mechanics are in app-boot; config fields are in the generated config catalog.

Application launch

Every supported Node application starts at the dsh CLI with a named profile. The shipped applications are dsh web (the deliberate alias for --profile web), dsh --profile headless, dsh --profile sdk, dsh --profile sdk-minimal, and dsh --profile acp. The TypeScript SDK resolves its same-version dsh dependency and selects sdk; custom plugin composition remains a profile plus ordered patch files, not another executable or inline application tree. sdk-minimal is a repository-owned standalone bundle behind the same launcher, not a caller-supplied Cordis tree.

Vendored CLIs, build-only and test-only executables, direct in-process plugin mounting, and the private browser WebWorker preview are not Harness application launchers. verify-application-entrypoints keeps every package bin, executable source, and root demo in an explicit class and rejects a Node application path that bypasses dsh.

The Python SDK follows the same application architecture. Its runtime wheel packages the normal dsh CLI as deepseek-harness-sdk-runtime-<platform>-<arch>, and the client launches dsh --profile sdk with an explicit Harness home by default. The minimal example selects the shipped sdk-minimal profile. Python exposes profile selection and ordered patch files rather than a complete Cordis tree; persistent external plugins are installed through dsh plugin. The removed private direct-config carrier has no compatibility bin or fallback parser.

Core packages

Here are some core packages that contribute to the Cordis tree.

Package	Owns	ctx key
core/session	The append-only SessionEvent log and in-memory store	ctx.sessions
core/system-prompt	Prompt-section and tool-schema assembly	ctx.systemPrompt
core/tools	The scoped tool registry and guarded execution pipeline	ctx.tools
core/agent	The Agent interface, live registry, and agent/* events	ctx.agents
core/agent-loop	The default driver implementing that interface	ctx.agentLoop
core/scope	The per-agent scoped-registration primitive	library, no key
llm/llm	Message and stream vocabulary plus the adapter seam	ctx.llm
webhook/webhook	Authenticated-delivery dispatch and Workspace Session creation	ctx.webhookRuntime

Events

Events are the extension points, and picking the right domain is the first decision in most changes.

• Session events are durable facts appended to the log and broadcast through session/event. Use one when the fact must survive a reload.
• Agent events (agent/*) carry a live Agent: inbox, step, status, request, validation, continuation. Use one to observe or intercept work in flight.
• Capability events attach policy and adapters to a seam (fs/*, tools/*, telemetry/*) without importing the loop.

The event map lists every event's producers and consumers.

Turn flow

A step is one model request plus the tools it calls. A turn is zero or more steps: it opens before its first input is claimed and closes once nothing is owed.

turn/start
  claim next-step input plus one queued message
  assemble prompt sections + tool schemas
  -> agent/pre-step                   reject | enter(messages, startsRequestSeries?)
     reject, or a first enter rewritten empty -> close the turn with no step
     step/start
     append entered messages as user/message
     derive model history from the log
     agent/request -> llm/stream -> agent/assistant-stream start
       agent/assistant-stream chunk*
       assistant/message | assistant/attempt -> agent/assistant-stream end
     tool/call* -> tools/pre-execute -> tools/execute -> tools/post-execute -> tool/result*
     step/end
     tools owe another request, or next-step input arrived -> claim -> next step
  -> agent/turn-stopping
turn/end

turn/*, step/*, user/message, assistant/message, assistant/attempt, and tool/* are durable session events; the rest are live extension points across three domains. agent/assistant-stream publishes process-local start, transient chunk, and end frames. The loop commits the complete compact stream as one message or log-only attempt before a committed end frame, and the Web Session-follow adapter is the live event's only remote consumer. agent/pre-step, agent/request, llm/stream, and the three tools/* events are waterfalls, whose listeners must call next() to delegate; agent/turn-stopping is serial and has no next().

Input reaches the driver through one inbox. Some messages wake it immediately; injected context waits in the inbox until another message does.

agent/pre-step decides what the model sees. Listeners may rewrite the claimed messages or reject them outright; a rejected or empty first claim still closes a durable turn that spent no step, so the log records the attempt. An enter decision may also set startsRequestSeries to begin a distinct model-message series: the loop then logs a fresh request/header (reason series, or change carrying startsSeries: true when the envelope changed too). A listener that rebuilds a downstream enter decision must spread it ({ ...decision, messages }) so the declaration survives. Each step reads the prompt sections and tool schemas that plugins registered.

Details: the sequence diagram, the tool pipeline, and cancellation and error recovery.

Session log

The session log is the source of the context the model sees. deriveMessages() projects model history from it. Each assistant/message embeds the exact compact timed stream that produced its assembled content; assistant/attempt retains settled failed, retried, cancelled, and stream-error attempts without adding model history. Fork, resume, transcripts, telemetry, and persistence all derive from these durable settlements, while live UI incrementality comes from agent/assistant-stream; a hard process loss before settlement leaves no durable attempt stream (decision).

Session consumers know only the current logical format. Header-only stat and list rescan each Session directory, select its numerically highest canonical generation, and translate a supported historical header without loading events or publishing a successor. A stored-session open selects that same generation, refuses a future version, or composes the static adjacent migration chain in memory, validates the final result, and exclusively publishes only that version-named successor beside the unchanged source before returning a handle; semantic interrupted-turn repair remains a handle consumer responsibility. JSONL v0 uses session.jsonl[.zstd], v1 and later use lowercase session.vN.jsonl[.zstd], and committed generation paths are never renamed, replaced, or deleted. The JSONL provider owns physical framing, compression, generation selection, and exclusive publication, while each adjacent migration package owns exactly one vN -> vN+1 step (decision).

Model-visible means logged. Anything that reaches a model request must be reconstructable from the log, and a runtime invariant asserts it. This is why a new model-visible input requires a new session event: extend SessionEventMap and render from the log.

Projection seam. dsh-session-projection owns ctx.sessionProjections: registered units fold committed events incrementally, host consumers read one typed state with stateOf(), and carriers batch cropped client views with snapshot(). A host reader either requires this service during activation or fails explicitly when the registry or required key is absent. Contributors may retain ctx.inject(['sessionProjections'], ...) registration without silently defaulting a missing host value. The agent loop registers shared turnBoundary state for its readers (decision).

Capability seams

A seam is a swappable capability with three roles: a Service Definition declaring the interface, a Service Provider implementing it, and a Consumer using it, commonly a model-facing tool. A package may combine roles, but one role alone is not a seam; adding a capability means designing all three (capability graph).

Seams are why one provider swap changes the whole product. Filesystem and subprocess providers share one execution world, so pointing them at a remote sandbox moves Bash, PTY, and LSP with them, with no provider forks. Subagent providers vary just as widely behind one interface, from a fresh child agent to a delegated turn in another product.

Experimental Agent Teams is a private opt-in coordination seam on ctx.agentTeams, with a durable roster, task board, and mailbox layered over continuable subagents.

Where new behavior goes

New behavior attaches to a documented extension point. Changing the loop itself updates this map.

Goal	Mechanism
Add a model provider	register its adapter on ctx.llm
Add a model-facing capability	register on ctx.tools; its schema joins prompt assembly
Give one session a different capability set	compose an agent preset; a service row there needs an isolate realm
Add shell execution	register a ctx.shell backend; the local one spawns through ctx.subprocess
Add persistent terminal execution	register a ctx.terminals backend plus dsh-tool-terminal
Add a human command	register on ctx.commands; it dispatches without a model turn
Add background work	register on ctx.jobs; job_* tools collect or stop it
Start a Session from an external webhook	register a trusted rule on ctx.webhookRuntime and mount a provider adapter
Add filesystem access or policy	register a ctx.fs provider or listen to fs/* events
Confine spawned processes	use a ctx.sandbox backend; consumers wrap argv before spawning
Intercept a request, tool, or turn	use its agent/* or tools/* event; agent/turn-stopping stops a turn
Add model-facing context	call agent.inject(); it lands in the next admitted request
Add UI or editor integration	drive ctx.agents and render from session/event
Add a Web Client Chat node	register a ConversationNodeDefinition + keyed renderer
Add durable session state	extend SessionEventMap; render and replay from the log
Generate session titles	register the sole ctx.sessionTitle provider
Manage a same-session objective	use ctx.goals; continue through agent/*
Fork a session at a turn boundary	ctx.agents.create({ sessionId, seed, meta: { parentSession, seedLength } }) — only agent-loop-published sessions persist
Store sessions in a new backend	implement SessionPersistence (create/open/stat/list/export) over the shared handle scaffolding
Scope a registration to one agent	use that agent's agent.ctx

The extension cookbook maps features to capabilities and indexes the step-by-step guides for packages, tools, LLM adapters, and settings cards. The Conversation subsystem owns Chat-node assembly.